Define a batch_size_per_proc to lower memory footprint, there's going to be a natural tradeoff

pmd.py

@@ -93,7 +93,9 @@ class DatasetsLoader(BaseLoader):
         self.dataset_name = dataset_name
         self.config_name = config_name
         self.num_proc = num_proc
-        self.batch_size = batch_size
+        # In order to not have memory explode we define a batch size per proc
+        self.batch_size_per_proc = batch_size // self.num_proc
+        assert self.batch_size_per_proc > 1
 
     @abstractmethod
     def cast_to_pmd_features(self, row: Dict) -> List[Dict[str, Any]]:
@@ -111,7 +113,7 @@ class DatasetsLoader(BaseLoader):
 
     def _generate_rows_casted_pmd_features(self, dset: Dataset, batch_start: int):
         dataset_size = len(dset)
-        batch_end = min(batch_start + self.batch_size, dataset_size)
+        batch_end = min(batch_start + self.batch_size_per_proc, dataset_size)
         batch = dset[batch_start:batch_end]
         rows = self.convert_batch_to_list_of_rows(batch)
 
@@ -120,14 +122,13 @@ class DatasetsLoader(BaseLoader):
             rows_casted_pmd_features += self.cast_to_pmd_features(row)
         return rows_casted_pmd_features
 
-
     def _generate_examples(self):
         dataset = load_dataset(self.dataset_name, self.config_name, split=self.split)
         dataset_size = len(dataset)
 
         # load batches and yield individual rows
         if self.num_proc == 1:
-            for batch_start in range(0, dataset_size, self.batch_size):
+            for batch_start in range(0, dataset_size, self.batch_size_per_proc):
                 rows_casted_pmd_features = self._generate_rows_casted_pmd_features(dset=dataset, batch_start=batch_start)
                 for row_casted_pmd_features in rows_casted_pmd_features:
                     yield row_casted_pmd_features
@@ -138,7 +139,7 @@ class DatasetsLoader(BaseLoader):
             with Pool(self.num_proc) as pool:
                 rows_casted_pmd_features = pool.imap(
                     partial(self._generate_rows_casted_pmd_features, dataset),
-                    range(0, dataset_size, self.batch_size),
+                    range(0, dataset_size, self.batch_size_per_proc),
                 )
                 for row_casted_pmd_features in rows_casted_pmd_features:
                     yield row_casted_pmd_features